Re: [RFC PATCH] diff: do not use creation-half of -B as a rename target candidate

Junio C Hamano [off-list ref] writes:

 * Here is what I am at the moment; I cannot quite explain (hence I
   cannot convince myself) why this is the right solution, but it
   seems to make the above sample case work without breaking any
   existing tests.  It is possible that the tests that would break
   without the "&& !p->score" bit are expecting wrong results, but I
   didn't look at them in detail.

Sadly, I think this is garbage.  "Do not consider creation-half of a
broken pair, ever" is too simple and cripples this case that starts
with two files A and B that are quite different:

	$ git add A B
	$ mv A B.new
        $ mv B A
        $ mv B.new B
        $ git diff -B -M

where the internal machinery breaks both A and B into these two file
pairs:

	delete A(old)
        create A(new)

	delete B(old)
        create B(new)

and then match them up to produce

	rename A to B
        rename B to A

The rule need to be "creation-half of a broken pair can be used as
the destination of a rename, if and only if its corresponding
deletion-half is used as the source of another rename elsewhere".
Under that condition, a file A that is completely rewritten to
become similar to another existing file B can be expressed as a
rename of B, because A is renamed away to make room in the same
change.

Fixing this is turning out to be more complex than I originally
hoped X-<.